search

cg1177 / DCAN

[AAAI 2022] DCAN: Improving Temporal Action Detection via Dual Context Aggregation
https://arxiv.org/abs/2112.03612
Apache License 2.0
18 stars 0 forks source link

Too low accuracy #4

Closed mugenggeng closed 1 year ago

mugenggeng commented 1 year ago

Excuse me, why did I get a low accuracy after training and testing according to your method

mugenggeng commented 1 year ago

mAP at tIoU 0.1 is 0.7184708874933091 mAP at tIoU 0.2 is 0.6910493253734883 mAP at tIoU 0.3 is 0.6523199517814122 mAP at tIoU 0.4 is 0.5964088928349324 mAP at tIoU 0.5 is 0.5201558475817378 mAP at tIoU 0.6 is 0.4196229250109158 mAP at tIoU 0.7 is 0.3085124834642902 mAP at tIoU 0.8 is 0.18811028051123077 mAP at tIoU 0.9 is 0.05531092062872859

mugenggeng commented 1 year ago

Excuse me, I would also like to ask whether AR@AN selects the first K proposals to calculate regression through setting the parameter max_avg_nr_proposals

cg1177 commented 1 year ago

Hello, I was working on some other projects before, so I just saw your issue.

Excuse me, why did I get a low accuracy after training and testing according to your method.

Unstable performance due to small dataset. You can try to run the training procedure multiple times.

Excuse me, I would also like to ask whether AR@AN selects the first K proposals to calculate regression through setting the parameter max_avg_nr_proposals.

The evaluation program is derived from the official, we have not modified it. You can view the evaluation code for details.